Otimização dos processos de inventário florestal por meio do aprendizado de máquina

• Ano de defesa: 2025
• Autor(a) principal: Ziembowicz, Matheus Morais
• Tipo de documento: Tese
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Federal de Santa Maria; Brasil; UFSM; Programa de Pós-Graduação em Engenharia Florestal; Centro de Ciências Rurais
• Palavras-chave em Português: Amostragem estratificada; Amostragem florestal de precisão; Eucalyptus x urograndis; Random Forest; Sensoriamento remoto; Stratified sampling; Precision forest sampling; Remote sensing; CNPQ::CIENCIAS AGRARIAS::RECURSOS FLORESTAIS E ENGENHARIA FLORESTAL
• Link de acesso: http://repositorio.ufsm.br/handle/1/38430

Planted eucalyptus forests form the backbone of the Brazilian forestry sector, with Eucalyptus × urograndis being the most widely used hybrid due to its rapid growth and high productivity. Forest inventory is an essential tool for sustainable management and production planning; however, conventional inventory methods often require high sampling intensities to ensure adequate levels of statistical precision, resulting in greater operational effort, higher costs, and longer fieldwork time. This scenario highlights a relevant methodological limitation in the execution of forest inventories, particularly in large-scale and structurally heterogeneous commercial plantations. In this context, the present study aimed to propose and evaluate a methodological approach capable of reducing the sampling intensity required in forest inventories of Eucalyptus × urograndis, while maintaining adequate levels of precision and statistical reliability, through the use of the Random Forest (RF) algorithm combined with vegetation indices derived from Sentinel-2 imagery. The choice of the RF algorithm is based on its robustness, ability to handle a large number of correlated variables, and consistent performance in recent forestry applications, especially when integrated with remote sensing (RS) data. The research was structured into two main stages. In the first stage, different vegetation indices were tested and evaluated using nonparametric statistical analyses (Kruskal–Wallis test and Spearman correlation), as well as RF performance metrics, including overall accuracy, the Kappa index, and cross-validation. This stage allowed the identification of the most relevant indices for structural differentiation of forest stands and validated the robustness of the algorithm in classifying productivity zones. In the second stage, the forest inventory was conducted using two sampling approaches: (i) simple random sampling and (ii) stratified sampling based on productivity zones generated by the RF model. The results were compared using descriptive statistics, analysis of variance (ANOVA), and an integrated synthesis of the required sampling intensity. The results demonstrated that the RF algorithm showed consistent performance in stand stratification, with high accuracy and Kappa values, as well as the formation of internally homogeneous and externally heterogeneous strata. The analysis of vegetation indices revealed significant correlations with productivity, confirming their relevance as auxiliary variables in the stratification process. When comparing inventory procedures, the stratified approach enabled average reductions between 60% and 74% in sampling intensity, while maintaining adequate precision levels across all evaluated stands. These findings confirm the central hypothesis of the thesis, demonstrating that the integration of machine learning and forest inventory represents an effective strategy to rationalize field effort and optimize sampling processes. It is concluded that the use of RF combined with vegetation indices derived from remote sensing not only provides statistical gains, but also represents a conceptual and practical advancement in forest planning, contributing to operational efficiency, sector sustainability, and the improvement of planted forest management in Brazil. Nevertheless, it is emphasized that the application of the proposed approach should consider the limitations associated with stands exhibiting greater structural heterogeneity, requiring a critical and site-specific implementation.